What is recency? Natural history of HIV

Lab assays to identify infection stages

This is great, but specificity and sensitivity are hard to control:

• individual level variation, confounding subtypes, ART status…

An alternative unit of time: genetic diversity

Case study: recency in Botswana

Ragonnet-Cronin et al. (2022) developed recency classifier

• end goal: characterizing undiagnosed cases in Botswana (Bhebhe et al. (2022))
• data: demographic data, some testing data and individual level NGS
• how: train ML model on known recent and known chronic infections, and use it to classify unknown.

Case study: recency in Botswana

HIV-phylo TSI: training data

sequences from PANGEA and BEEHIVE with known infection range.

HIV-phylo TSI: intrahost diversity

choice of predictors based on viral diversification following infection.

1. “Accumulated” mutations on codons.
2. “Depth” of host-specific subtree in phylogeny. (LRTT)

But why these measures of intrahost diversity?

HIV-phylo TSI performances: bias

HIV-phyloTSI performances: robustness to ART (?)

Next Generation Sequencing

• simultaneous sequencing of multiple viral particles to capture within-host diversity
• PANGEA protocols: I (2012) and II (2018)
• TSI pipeline starts with the *.bam, *_ref.fasta produced by shiver

Summary

graph LR
    subgraph S1["Primary Data Patients"]
        direction TB
        I01[(*.bam)] 
        I02[(*ref_fasta)] 
        I02b[(*BaseFreqs_WithHXB2.csv)]
    end

    subgraph S2["Reference Sequences"]
        I03[(*.fasta)] 
    end

    subgraph S3["Multiple Sequence Alignment"]
        P1(mafft)
        O1[(*.fasta)]
    end

    subgraph S4["Constructing Phylogenies"]
        direction TB
        P2(IQTREE)
        O2[( *.iqtree )]
    end
    
    subgraph S2b["Computing MAFs"]
      direction TB
      P2b(script)
      O2b[(maf.csv)]
    end

    subgraph S5["Analysing phylogenies"]
        direction TB
        P3( PhyloscannerR ) 
        O3[(*PatStats.csv)]
    end

    subgraph S6["Obtain TSI estimates"]
        direction TB
        P4( HIV-phyloTSI ) 
        O4[(TSI.csv)]
    end
    
    classDef Red fill:#F84E68
    class S3,S4 Red
    classDef Orange fill:orange
    class S5,S6 Orange

I02 --> P1
I01 --> P1 
I02b --> P2b --> O2b --> P4
I03 --> P1
    P1 --> O1
O1 --> P2 
    P2 --> O2
O2 --> P3 
    P3 --> O3
O3 --> P4 
    P4 --> O4

Installing dependencies

The ‘ingredients’ to run the analyses are: the ML algorithm; the data and the code dependencies. The below code chunk allows you to download everything that is needed for the analyses.

# change directory to where you want to install HIV-phyloTSI repo
cd $HOME/git            # this is where I install git packages 
# cd $HOME && mkdir git && cd git

# clone directories necessary to run the analysis
# BDIs code and workshop materials
git clone git@github.com/BDI-pathogens/HIV-phyloTSI.git
git clone git@github.com/abriz97/HIV-phyloTSI-workshops.git
# store paths to 2 above directories
DIR_WORKSHOP="$(pwd)/HIV-phyloTSI-workshops"
DIR_PROGRAM= "$(pwd)/HIV-phyloTSI"

# install python dependencies for HIV-phyloTSI and load the environment
conda env create -f HIV-phyloTSI-workshops/hivphylotsi.yml
conda activate hivphylotsi

Running the algorithm

Once all the ingredients are there, we can start cooking. It is relatively simple to run the analyses, even though we need to be precise in the way we specify the paths to the input data.

# Run HIV-phyloTSI on input data.
python $DIR_PROGRAM/HIV-phyloTSI.py \
    -d $DIR_PROGRAM/Model \
    -p $DIR_WORKSHOP/HIV-phyloTSI-workshops/input/ptyr1_patStats.csv \ 
    -m $DIR_WORKSHOP/HIV-phyloTSI-workshops/input/phsc_input_samples_maf.csv \ 
    -o $DIR_WORKSHOP/HIV-phyloTSI-workshops/output/ptyr1_tsi_workshop.csv
  
# print header of output to make sure it exists:
head $DIR_WORKSHOP/HIV-phyloTSI-workshops/output/ptyr1_tsi_workshop.csv

Predictors used

dtsi  <- fread(file.path(git_root, "data/ptyr1_tsi.csv"))
plot_predictors_from_tsi_output(dtsi, exclude="dual")

Evaluation on seroconverters

dall  <- merge(dtsi, ddates)
plot_cross_interval_tsisero(dall, sqroot=TRUE)

Histogram of estimates TSIs

p_hist <- plot_histogram_tsi(dtsi)

Prevalence of recent infections

where \(MDRI\) = mean duration of recency of infection.

• \(MDRI = P[\text{ever being classified as recent}] \times E[ TSI | \text{classified as recent} ]\)
• not a constant: viral suppression, underlying population.

Here we focus on the numerator.

Simulation study:

In simulated settings, the prevalence of recent infections is well estimated, despite at least 2% of recent classifications are wrong.

Application in Zambia

When directly applied to real-world data, HIV-phyloTSI generally produces smaller TSI
for known recent infections as compared to people with unknown first positive date:

=> Can be used to compare median TSIs among population subgroups.

Suggestions on TSI comparison among groups

Comparing simple summary statistics can be misleading:

• need to make sure statistic is reliable
• need to chose summary statistic: median

For reliability:

1. make sure groups are large enough (minimum of 40 per group)
2. Perform bootstrap to estimate uncertainty around

Suggestions on TSI comparison: bootstrap

Bootstrapping: statistical technique which recycles analysis data to estimate uncertainty around an estimator (e.g. median).

1. Synthetic data sets of the same size are obtained by sampling with replacement from the data.
2. The estimator is computed on each synthetic data set.
3. The distribution of the estimators obtained on the syntetic data sets is summarised through 95% quantiles.

Other analyses on population level outputs:

Estimation of generation time distribution

Enriching source-recipient pairs by providing time since infection.

• Zambia (Hall et al. 2021)
• Rakai: (Monod et al. 2023)

Summing up

• HIV-phyloTSI is a novel algorithm to estimate infection dates.

• alternative to serological assays which allows more control on definition of recency (Robust to subtype and ART usage)

• preliminary analyses and simulation studies demonstrate good performances at population level

• but may be inaccurate at the individual level: very large prediction intervals.

• Can help us explore answers to unanswered questions